有冇巴打可以講吓back propagation,k-nearest neighbor,support vector machine同decision tree,我ee仔冇ml background但依幾日要知佢做緊乜
Backpropagation即係一個Algorithm去計 neural network入面parameter，原理係用network output result個error 去層層逆推計parameter以達致最適error，行完結果係得到一個neural network，我諗做分類
k-nearest neighbor (K-means)做data分類，原理係每粒data附近k粒最近既data應該大部分係同類，行完結果係一堆分好類既data
Support vector machine (SVM)做data分類，原理係用data投影喺條界上既位置計距離，行完結果係得到一條界cut係兩邊data既正中間
Decision tree我諗指分類 learning嗰隻，原理係逐個feature起個條件分類，得到一個跟條件既分類法，例子如下: 你身高夠唔夠180cm? 夠>你會考有無30分? 有>你賓周有無30cm長? 有>...諸如此類...>你係高登仔
利申: 其實我唔太識，係上網自學，我都無background，不過見好似無人答你先講兩句，有咩唔明大家研究下，下面係一啲舊年上coursera堂送既reference
Backpropagation Algorithm
http://ufldl.stanford.edu/wiki/index.php/Backpropagation_Algorithm
An Idiot’s guide to Support vector machines (SVMs)
http://web.mit.edu/6.034/wwwbob/svm-notes-long-08.pdf
Bp講緊由錯誤中學習 用自定既error measure(最簡單就係euclidean distance)去improve返neural network, 如果高既learning rate就會學曬呢一次既野 而冇左好多以前學左既野
本意係minimize error, geometrically 睇就係要去到error curve既minimum point, 所以涉及patial d去搵呢個pt, 好似dse curve sketching搵minimum pt咁
如果你有d numerical analysis底就可以理解為step size太大 去唔到optimal point
As a signal黎講就係sample得太疏 唔會知道真正既最低點（唔係100%岩 只係類近d既講法）
K-mean多數係做clustering 將最近(呢個distance又係要自己define, 多數係euclidean)既痴埋一齊
SVM多數係用黎做classification 要睇你本身用咩kernel, 冇就係linear, 搵一條optimal既直線分開兩個group
係喎，我係咪都叫分類，講得唔清楚
分類有分有監督同無監督，有監即係手頭上既sample一早已經知邊個打邊類，用現有預估將來，無監即係而家唔知邊個打邊個，靠估分返一群群
clustering中文叫聚類，classification中文叫分類，前者係unsupervised無監督，後者係supervised有監督。k-means係unsupervised做clustering，k-nearest neighbors係supervised做classification，應該係咁樣
k-nearest neighbors唔係同 k-means一樣咩? 爭個supervised vs unsupervised?
巴打禮拜六都起到身，我都想乘機問下euclidean distance係咪即係L2 norm? 一時一個名好混亂，幾時用邊個名?
